Activity time: Saturday, January 9, 2021 14:00-18:00
Venue: Floor 1, Block E, Wangjing Science and Technology Park, Chaoyang District, Beijing
Event registration: poke me to sign up
| Activity Introduction
Unmanned vehicles and unmanned aerial vehicles will have a profound impact on our lives in the future. As a leading company in the life service industry, Meituan has been committed to cutting-edge exploration in the field of unmanned driving in recent years. Meituan's unmanned distribution team has been deeply involved in urban distribution scenarios, and its research directions cover many fields such as robot spatial positioning, instance segmentation, trajectory prediction, and planning control. In this technical salon, we will share with you the latest achievements of the unmanned distribution team in these directions this year.
Seller
Dr. Mao Yiian| Head of UAV Business Department, Meituan
Responsible for the business exploration of Meituan UAV in the city's low-altitude logistics distribution. Joined Meituan in December 2018. Prior to this, he was the founder and CEO of Beijing Erlangshen Technology Company, bringing environmental awareness and autonomous planning capabilities to drones. In 2018, he was named as Beijing Haiju Engineering Expert, Beijing Chaoyang District Phoenix Project Technical Expert.
He received a bachelor's degree in electronic engineering from Tsinghua University in 2001 and a doctorate degree in electronic engineering from the University of Maryland in 2006. He once worked at Qualcomm Research Institute in the United States, leading and participating in the development and commercialization of important projects such as MPEG DASH. Won the Qualcomm Research RoCStar Award in 2014. He is the inventor of more than 30 international patents, has published more than 20 international journals and conference articles, and his papers and patents have been cited thousands of times.
| Event Calendar Poster
| Share Introduction
Topic 1: Introduction to CenterMask, a one-stage instance segmentation algorithm based on point representation
Wang Yuqing | Meituan Algorithm Engineer
Graduated from Nankai University with a master's degree and joined Meituan in 2019. Currently responsible for target detection and instance segmentation related work in the application of high-precision map element extraction.
Introduction: Introduce the paper published on CVPR2020 by the high-precision map group of the unmanned vehicle distribution center of Meituan AI platform, based on the point representation of the one-stage instance segmentation algorithm CenterMask.
Essay topic:
CenterMask: single shot instance segmentation with point representation
Topic 2: Visual positioning algorithm based on geometric constraints of 3D scene
Tian Mi | Meituan Algorithm Engineer
An algorithm engineer at Meituan's unmanned vehicle distribution center. He joined Meituan in 2018 and is currently responsible for the research and development of visual positioning algorithms in the unmanned vehicle business.
Introduction: Introduce the paper published on ICRA 2020 by the visual positioning group of the unmanned vehicle distribution center of the Meituan AI platform, based on the visual positioning algorithm of 3D scene geometric constraints.
Essay topic:
3D Scene Geometry-Aware Constraint for Camera Localization with Deep Learning
Topic 3: Multi-agent trajectory prediction algorithm based on global interaction
Zhu Yanliang | Meituan Algorithm Engineer
An algorithm engineer at Meituan's unmanned vehicle distribution center. He joined Meituan in 2018 and is currently responsible for the analysis and prediction of obstacles in the unmanned vehicle business.
Introduction: Introduce how the pnc group of Meituan unmanned vehicle distribution center participates in the ICRA 2020 trajectory prediction competition, and the multi-agent trajectory prediction algorithm based on global interaction.
Entry method name:
Robust Trajectory Forecasting for Multiple Intelligent Agents in Dynamic Scene
Topic 4: A Pedestrian Trajectory Prediction Network Model Based on Spatio-temporal Graph Convolution with Interactive Perception
Fan Mingyu | Meituan Researcher
R&D consultant and researcher of Meituan Autonomous Vehicle Distribution Center. He joined Meituan in 2019 and is currently mainly engaged in research and development of trajectory prediction and path planning.
Content introduction: This article proposes an interactive perception spatiotemporal graph convolution model method based on attention model, which can realize dynamic and adaptive obstacle interaction spatiotemporal feature modeling to capture complex interactive information between multiple obstacles.
Essay topic:
An Attention-based Interaction-aware Spatio-temporal Graph Neural Network for Trajectory Prediction
Topic 5: Implementation of the online calibration algorithm of binocular external parameters on VIO
Lang Xiaoming | Meituan Algorithm Engineer
An algorithm engineer of the UAV Business Department of Meituan, joined Meituan in 2019 and is currently responsible for the research and development of UAV visual positioning algorithm.
Introduction: Introduce the paper on the real-time calibration of binocular external parameters by the visual positioning group of Meituan UAV Distribution Center on ICRA2020.
Essay topic:
Stereo Visual Inertial Odometry with Online Baseline Calibration
Topic 6: Self-supervised depth-position joint learning for dynamic scenes
Gao Feng | Master student, Tsinghua University
Obtained a bachelor's degree from the Department of Electronics of Tsinghua University in 2019, and is currently studying for a master's degree in the Energy-Efficient Computing Group of the Nano-Integrated Circuit and System Laboratory of the Department of Electronics. His supervisor is Professor Wang Yu. Research interests include robot positioning and navigation, reinforcement learning, and multi-machine collaborative intelligence.
Content introduction: Monocular depth estimation and visual odometry play a very important role in fields such as autonomous driving. In recent years, the self-supervised learning method of simultaneously learning depth and camera pose from unlabeled videos has gained great attention. However, in dynamic scenes, the existence of moving objects can interfere with the generation of self-supervised signals. In response to this problem, we propose that we can learn the motion of the scene and the motion of the camera at the same time, and adaptively generate supervision signals. Related work was published in the CoRL2020 conference.
Essay topic:
| Registration method
Recognize the QR code and register for free